Senior AI/ML Engineer

Description

This role is also open to candidates living in the EU or EEA. We can't consider applications from other locations.

If you've built something before, start here

Most ML jobs hand you a benchmark. This one hands you a business problem with paying customers and their real data. Construction tenders are long, messy and full of trade jargon, and every decision a customer makes on our platform becomes a labeled example you can train on.

We need someone who can turn that data into models that customers notice: from the first experiment to the model serving production traffic, and back again when a customer says it got something wrong. If you've founded a company (whether it worked out or not), been the first ML engineer somewhere, or shipped a model people paid for, this role was written for you.

About Scalera

Scalera is an ETH Zurich spin-off building AI for the €200B+ European construction industry. Manufacturers, wholesalers and suppliers in Switzerland, Austria and Germany still run their most document-heavy work by hand, from reading 300-page tenders to retyping purchase orders into their ERP. We automate that work one process at a time, and ML sits at the center of every one of them.

What you'll own

You'll own the ML behind our products, end to end. In practice that means:

Training and fine-tuning : embedding models, rerankers, classifiers and LLMs, trained on our own data.

Document understanding : models that pull structure, quantities and product specs out of tenders, orders and catalogs.

Product matching : retrieval and LLM agents that match tender positions to the right products.

Evaluation : offline benchmarks, per-customer evals and gates that keep a new model from shipping if it makes things worse.

Training and serving infrastructure : automated retrain pipelines, GPU serving and vector search in production.

Customer problems : when a model gets something wrong for a customer, finding out why and fixing it.

What success looks like after 3 months

There's a model or part of the pipeline people now think of as yours, and you're the one deciding what happens to it next.

You've spotted a problem nobody asked you to solve, made the case for it, and shipped a first version.

A model you trained runs in production, and customers can see the difference in their results.

Retraining runs automatically, and nothing reaches customers without passing an eval you designed.

You've cut LLM cost or latency by replacing an expensive step with a smaller model of your own.

You've decided at least one thing we should stop building.

You'll probably thrive here if

You've taken something from zero to one: a company, a model in production, or a side project with real users.

You've trained and fine-tuned models yourself, ideally embeddings, retrieval or document models, and evaluated them honestly on held-out data.

You've built with LLMs in production and know where agents fail.

You write strong Python and can run what you build: Docker, CI/CD, and cloud on GCP, AWS or Azure.

You measure before you believe, and a single aggregate score doesn't convince you.

You'd rather open a customer's actual document than wait for a spec.

You're comfortable with ambiguity. You can make progress while the problem is still half-defined, which at a startup it usually is.

You want a large scope of responsibility and are happy to own things well beyond your job title.

Rapidly changing requirements don't throw you. When customers or priorities shift, you adapt and keep going.

You enjoy direct customer contact and want to be on customer calls.

You write clearly in English. German is a plus, not a requirement.

You probably won't enjoy it if

You just want to code and go home.

You don't like talking to customers.

Ambiguity and chaos bother you.

You want a well-defined research agenda handed to you.

You'd rather hand a model to someone else to deploy and debug.

Your main goal is managing people soon. Here, scope grows with what you own, not with headcount.

Our stack

Python and PyTorch, open-weight and frontier LLMs, LangGraph, Langfuse, PostgreSQL with pgvector, Pinecone, and GCP. You don't need to know all of it on day one.